fix(quic): tier-local round-robin so priority survives every drain

The previous priority-then-round-robin shape applied the rotating
start-index globally over the sorted list, so the cross-tier order
flipped on alternate drains: drain N had high-priority first, drain
N+1 advanced streamRoundRobinStart and had low-priority first. The
priority hint was silently defeated under any sustained traffic.

Replace it with strict priority across tiers + round-robin only
within each same-priority tier. Higher tiers always drain ahead of
lower ones; same-priority peers continue to take turns via the
existing rotating start. Default-priority callers see no behaviour
change (single tier, identical rotation semantics).

Tighten the test: drain twice and assert the higher-priority stream
emits first on BOTH drains — the regression case that the single-
drain version of the test missed. Add a regression guard for the
same-priority round-robin so a future refactor can't silently
serialise on the first stream.

https://claude.ai/code/session_01KWdr4RjVvyYZfEuPVaQfUa
This commit is contained in:
Claude
2026-05-06 20:34:39 +00:00
parent f1034b1f53
commit 72295915de
2 changed files with 132 additions and 80 deletions
@@ -410,58 +410,74 @@ private fun buildApplicationPacket(
// insertion-ordered and stays in sync with the streams map.
val streamsView = conn.streamsListLocked()
if (streamsView.isNotEmpty()) {
// Priority-then-round-robin: stable sortedByDescending preserves
// insertion order within a tier, so same-priority streams keep
// the rotating start-index round-robin behaviour. Higher-priority
// streams (e.g. moq-lite newer-sequence group streams) drain
// first under congestion. Default priority is 0; if every stream
// is at the default, iteration order matches pre-priority code.
// Cost: O(N log N) per drain pass and one transient list
// allocation. N is small (110 in the moq-lite audio path); if
// it ever grows enough to matter, switch to an indirect index
// sort or maintain an incrementally-sorted view on setPriority.
// Strict priority across tiers, round-robin within each tier.
// Higher-priority streams (e.g. moq-lite newer-sequence group
// streams) ALWAYS drain ahead of lower-priority ones; the
// rotating start-index only rotates among same-priority peers.
// This is the spec-aligned shape — applying the rotation
// globally over the sorted list would flip cross-tier order on
// alternating drains, defeating the priority hint entirely.
//
// Default priority is 0; if every stream is at the default, all
// streams form a single tier and iteration order matches the
// pre-priority round-robin behaviour exactly.
//
// Cost: O(N log N) per drain pass plus one transient sorted
// list. N is small (110 in the moq-lite audio path); if it
// ever grows enough to matter, switch to an indirect index sort
// or maintain an incrementally-sorted view on setPriority.
val sorted =
if (streamsView.size > 1) streamsView.sortedByDescending { it.priority } else streamsView
val start = conn.streamRoundRobinStart % sorted.size
for (i in sorted.indices) {
if (packetBudget <= 64) break
val stream = sorted[(start + i) % sorted.size]
val streamRemaining = (stream.sendCredit - stream.send.sentOffset).coerceAtLeast(0L)
// Skip if both stream and connection have no credit; FIN-only
// (zero-byte) chunks may still go through because they don't
// consume credit.
if (streamRemaining <= 0L && connBudget <= 0L && !stream.send.finPending) continue
val effectiveCap = minOf(streamRemaining, connBudget)
val maxBytes =
minOf(packetBudget - 32, effectiveCap.coerceAtMost(Int.MAX_VALUE.toLong()).toInt())
val chunk = stream.send.takeChunk(maxBytes = maxBytes) ?: continue
if (chunk.data.isNotEmpty() || chunk.fin) {
frames +=
StreamFrame(
streamId = stream.streamId,
offset = chunk.offset,
data = chunk.data,
fin = chunk.fin,
explicitLength = true,
)
// Step C of the deferred-follow-ups pass: track this
// STREAM emission so RFC 9002 retransmit can re-queue
// the byte range on loss. SendBuffer.markLost (commit B)
// moves the range from in-flight back to the retransmit
// queue, and the next takeChunk replays it.
tokens +=
RecoveryToken.Stream(
streamId = stream.streamId,
offset = chunk.offset,
length = chunk.data.size.toLong(),
fin = chunk.fin,
)
packetBudget -= chunk.data.size + 32
connBudget -= chunk.data.size
conn.sendConnectionFlowConsumed += chunk.data.size
val rotation = conn.streamRoundRobinStart
var tierStart = 0
outer@ while (tierStart < sorted.size) {
// Walk the contiguous run of same-priority streams.
val tierPriority = sorted[tierStart].priority
var tierEnd = tierStart + 1
while (tierEnd < sorted.size && sorted[tierEnd].priority == tierPriority) tierEnd++
val tierSize = tierEnd - tierStart
val tierRotation = if (tierSize > 1) rotation % tierSize else 0
for (k in 0 until tierSize) {
if (packetBudget <= 64) break@outer
val stream = sorted[tierStart + ((tierRotation + k) % tierSize)]
val streamRemaining = (stream.sendCredit - stream.send.sentOffset).coerceAtLeast(0L)
// Skip if both stream and connection have no credit; FIN-only
// (zero-byte) chunks may still go through because they don't
// consume credit.
if (streamRemaining <= 0L && connBudget <= 0L && !stream.send.finPending) continue
val effectiveCap = minOf(streamRemaining, connBudget)
val maxBytes =
minOf(packetBudget - 32, effectiveCap.coerceAtMost(Int.MAX_VALUE.toLong()).toInt())
val chunk = stream.send.takeChunk(maxBytes = maxBytes) ?: continue
if (chunk.data.isNotEmpty() || chunk.fin) {
frames +=
StreamFrame(
streamId = stream.streamId,
offset = chunk.offset,
data = chunk.data,
fin = chunk.fin,
explicitLength = true,
)
// Step C of the deferred-follow-ups pass: track this
// STREAM emission so RFC 9002 retransmit can re-queue
// the byte range on loss. SendBuffer.markLost (commit B)
// moves the range from in-flight back to the retransmit
// queue, and the next takeChunk replays it.
tokens +=
RecoveryToken.Stream(
streamId = stream.streamId,
offset = chunk.offset,
length = chunk.data.size.toLong(),
fin = chunk.fin,
)
packetBudget -= chunk.data.size + 32
connBudget -= chunk.data.size
conn.sendConnectionFlowConsumed += chunk.data.size
}
}
tierStart = tierEnd
}
conn.streamRoundRobinStart = (start + 1) % streamsView.size
conn.streamRoundRobinStart = (rotation + 1) % streamsView.size
}
if (frames.isEmpty()) return null
@@ -112,45 +112,81 @@ class QuicConnectionWriterTest {
@Test
fun writer_drains_higher_priority_streams_before_lower_priority() {
// T11.3 follow-up: the writer's drain loop must iterate streams
// in descending priority order so moq-lite group streams with a
// higher sequence number drain ahead of older ones under
// congestion. Pre-fix this test, the writer iterated in stable
// round-robin order regardless of priority — so a backlog of
// retransmits on an older group could starve the listener of
// fresh frames. We pin the load-bearing invariant by inspecting
// StreamFrame order in a single emitted packet: low-priority
// stream is opened FIRST (so insertion order would normally win
// round-robin), but the high-priority stream's bytes must land
// earlier in the packet.
// T11.3 follow-up: priority must dominate iteration order on
// EVERY drain, not just the first. The naive "sort by priority
// then apply the existing rotating start globally" shape looks
// right at a glance but the rotating start advances on every
// drain, so cross-tier ordering flips on alternate drains and
// the priority hint is silently defeated. The correct shape is
// strict priority across tiers + round-robin only within a
// tier, which we pin here by draining TWICE and asserting the
// higher-priority stream's StreamFrame lands first in BOTH
// packets. Low-priority stream is opened FIRST so insertion-
// order can't accidentally pass for priority ordering.
runBlocking {
val (client, pipe) = connectedClient()
val low = client.openBidiStream()
val high = client.openBidiStream()
low.priority = 0
high.priority = 10
// Distinct payloads small enough to coexist in one packet.
val lowPayload = ByteArray(200) { 0xAA.toByte() }
val highPayload = ByteArray(200) { 0xBB.toByte() }
low.send.enqueue(lowPayload)
high.send.enqueue(highPayload)
val datagram = drainOutbound(client, nowMillis = 0L)
assertNotNull(datagram, "drain must emit a packet")
val frames = pipe.decryptClientApplicationFrames(datagram)
assertNotNull(frames, "decrypt must succeed at the application level")
val streamFrames = frames.filterIsInstance<StreamFrame>()
assertEquals(
2,
streamFrames.size,
"expected one StreamFrame per stream in this drain, got $streamFrames",
)
assertEquals(
high.streamId,
streamFrames[0].streamId,
"higher-priority stream must drain first; saw ${streamFrames.map { it.streamId }}",
)
assertEquals(low.streamId, streamFrames[1].streamId)
repeat(2) { round ->
low.send.enqueue(ByteArray(200) { 0xAA.toByte() })
high.send.enqueue(ByteArray(200) { 0xBB.toByte() })
val datagram = drainOutbound(client, nowMillis = 0L)
assertNotNull(datagram, "drain on round $round must emit a packet")
val frames = pipe.decryptClientApplicationFrames(datagram)
assertNotNull(frames, "decrypt must succeed on round $round")
val streamFrames = frames.filterIsInstance<StreamFrame>()
assertEquals(
2,
streamFrames.size,
"expected one StreamFrame per stream on round $round, got $streamFrames",
)
assertEquals(
high.streamId,
streamFrames[0].streamId,
"higher-priority stream must drain first on round $round; " +
"saw ${streamFrames.map { it.streamId }}",
)
assertEquals(low.streamId, streamFrames[1].streamId, "low second on round $round")
}
}
}
@Test
fun writer_round_robins_within_a_priority_tier() {
// Regression guard for the tier-local round-robin: same-priority
// streams must still rotate so an early-opened stream doesn't
// monopolise a packet's stream-frame slot indefinitely. We open
// three streams at the default (0) priority and verify the
// rotating start advances by one per drain, matching the
// pre-priority behaviour.
runBlocking {
val (client, pipe) = connectedClient()
val a = client.openBidiStream()
val b = client.openBidiStream()
val c = client.openBidiStream()
// All default priority — single tier, three streams.
val expectedRotation =
listOf(
listOf(a.streamId, b.streamId, c.streamId),
listOf(b.streamId, c.streamId, a.streamId),
listOf(c.streamId, a.streamId, b.streamId),
)
for ((round, expected) in expectedRotation.withIndex()) {
a.send.enqueue(ByteArray(64) { 0xA1.toByte() })
b.send.enqueue(ByteArray(64) { 0xB2.toByte() })
c.send.enqueue(ByteArray(64) { 0xC3.toByte() })
val datagram = drainOutbound(client, nowMillis = 0L)
assertNotNull(datagram, "drain $round must emit a packet")
val frames = pipe.decryptClientApplicationFrames(datagram)
assertNotNull(frames, "decrypt must succeed on round $round")
val ids = frames.filterIsInstance<StreamFrame>().map { it.streamId }
assertEquals(expected, ids, "round $round round-robin order")
}
}
}